Enhanced BLAS/LAPACK return code interpretation and logging by ChrisRackauckas-Claude · Pull Request #739 · SciML/LinearSolve.jl

ChrisRackauckas-Claude · 2025-08-18T23:41:59Z

Summary

This PR enhances the logging system introduced in #622 by adding detailed BLAS/LAPACK return code interpretation and extended logging information.

Changes

1. BLAS/LAPACK Return Code Interpretation

Added interpret_blas_code() function that provides human-readable explanations for all BLAS/LAPACK info codes
Categorizes errors (singular_matrix, not_positive_definite, convergence_failure, etc.)
Operation-specific interpretations for different LAPACK functions (getrf, potrf, geqrf, gesdd, etc.)

2. Extended Logging Information

Matrix properties (size, type, condition number when feasible)
Memory usage estimates for BLAS operations
Performance timing metrics with opt-in verbosity control
Additional context for debugging numerical issues

3. New Verbosity Controls

Added five new verbosity settings to the LinearVerbosity system:

blas_errors: Controls BLAS/LAPACK error messages (default: Warn)
blas_info: Controls informational messages (default: None)
blas_success: Controls success messages (default: None)
blas_invalid_args: Controls invalid argument errors (default: Error)
blas_timing: Controls performance timing (default: None)

4. Integration Example (BLISLUFactorization)

Demonstrated integration with the BLIS extension to show how the new logging can be applied to existing solvers.

5. Documentation

Updated verbosity documentation with all new BLAS options
Added comprehensive section on BLAS/LAPACK return codes
Included examples demonstrating the enhanced logging capabilities

6. Tests

Added comprehensive test suite for BLAS return code interpretation
Tests for different error categories
Verbosity integration tests

Example Output

When a singular matrix is encountered:

BLAS/LAPACK dgetrf: Matrix is singular
  Details: U(2,2) is exactly zero. The factorization has been completed, but U is singular
  Return code (info): 2
  matrix_size: (2, 2)
  element_type: Float64
  condition_number: Inf
  memory_usage_MB: 0.0

Test Plan

Tests pass locally
Documentation updated
Backwards compatible (all new features are opt-in)

🤖 Generated with Claude Code

This commit builds upon PR SciML#622's verbosity system by adding: 1. Detailed BLAS/LAPACK return code interpretation - Human-readable explanations for all BLAS/LAPACK info codes - Categorized errors (singular_matrix, not_positive_definite, etc.) - Operation-specific interpretations for getrf, potrf, geqrf, etc. 2. Extended logging information for BLAS operations - Matrix properties (size, type, condition number) - Memory usage estimates - Performance timing metrics - Contextual information for debugging 3. New verbosity controls - blas_errors: Controls BLAS/LAPACK error messages (default: Warn) - blas_info: Controls informational messages (default: None) - blas_success: Controls success messages (default: None) - blas_invalid_args: Controls invalid argument errors (default: Error) - blas_timing: Controls performance timing (default: None) 4. Integration with BLISLUFactorization - Added detailed logging to the BLIS extension - Includes timing and error interpretation 5. Comprehensive documentation - Updated verbosity documentation with new BLAS options - Added section on BLAS/LAPACK return codes - Examples demonstrating enhanced logging capabilities 6. Tests - Added test suite for BLAS return code interpretation - Tests for different error categories - Verbosity integration tests This enhancement makes debugging numerical issues much easier by providing clear, actionable information when BLAS/LAPACK operations encounter problems. 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <noreply@anthropic.com>

Based on review feedback: 1. Removed all BLAS timing functionality - Removed time_blas_operation function - Removed blas_timing verbosity setting - Removed timing from documentation and tests 2. Made get_blas_operation_info conditional on logging - Only calls get_blas_operation_info when logging is actually enabled - Avoids unnecessary computation when logging is disabled 3. Made condition number computation optional - Added compute_condition parameter (default: false) - Only computes condition number when explicitly requested - Avoids expensive computation for large matrices 4. Updated BLISLUFactorization extension - Removed timing calls - Only gathers operation info when needed for logging - More efficient checking of verbosity levels 5. Updated documentation - Removed all references to timing functionality - Documented optional condition number computation 6. Updated tests - Removed timing tests - Added tests for optional condition number computation These changes make the logging more efficient by avoiding unnecessary computations when logging is disabled. 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <noreply@anthropic.com>

@warn

Based on review feedback: 1. Replaced direct logging with @SciMLMessage macro - Updated log_blas_info to use @SciMLMessage with proper syntax - Structured log messages with appropriate categories and groups - Removed direct @info/@warn calls 2. Added condition_number as a separate verbosity setting - Added condition_number to LinearNumericalVerbosity struct - Condition number computation now controlled by verbosity setting - Default is Verbosity.None() to avoid expensive computation 3. Fixed verbosity checks - Use verbosity_to_int() for proper verbosity level checking - Ensures correct comparison with Verbosity.None() 4. Updated BLISLUFactorization extension - Uses @SciMLMessage for success messages - Proper verbosity checks before computing operation info - Only calls get_blas_operation_info when logging is enabled 5. Updated documentation - Added condition_number verbosity setting documentation - Clarified that condition number is controlled by verbosity 6. Updated tests - Tests now use verbosity parameter for condition number control - Verifies condition number is not computed when disabled These changes ensure efficient logging with proper use of the SciMLLogging infrastructure and make condition number computation a controllable option. 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <noreply@anthropic.com>

Replaced the last remaining @info call in check_and_log_lapack_result with @SciMLMessage macro for consistency with the SciMLLogging system. 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <noreply@anthropic.com>

Replace equality checks (!=) with identity checks (!==) for Verbosity.None() comparisons to enable compile-time optimization. This ensures logging code is completely compiled out when disabled, providing zero-cost abstraction for the verbosity system. 🤖 Generated with [Claude Code](https://claude.ai/code) Co-Authored-By: Claude <noreply@anthropic.com>

ChrisRackauckas · 2025-08-20T20:38:13Z

@jClugstor how do we set these up with verbosity levels?

jClugstor · 2025-08-20T22:17:10Z

You're talking about distinct levels like, level 1 should give some info, level 2 should give some more etc. ?

We could add Verbosity types like Verbosity.Low(), Verbosity.Medium(), Verbosity.High() to SciMLLogging, and make it so that the constructors for LinearVerbosity take them and look at a dictionary to see what each setting should be set to.

ChrisRackauckas · 2025-08-21T18:56:48Z

Yeah that would be a good thing to do, with some definition of what those mean.

jClugstor

Seems like everything is consistent with how SciMLLogging and the LinearVerbosity stuff works.

jClugstor · 2025-08-21T21:02:01Z

+        prob = LinearProblem(A, b)
+
+        # This should fail due to singularity but not throw
+        sol = solve(prob, LUFactorization(); verbose=verbose)


This doesn't actually test that a message was logged at the correct level.
It might be good to have sol = @test_logs (:warn, "BLAS/LAPACK dgetrf: Matrix is singular Details: U(2,2) is exactly zero. The factorization has been completed, but U is singular" match_mode=:any solve(prob, LUFactorization(); verbose=verbose) with whatever the actual log message ends up being, so that the tests make sure the correct things end up in the logs.

give the suggestion?

jClugstor · 2025-08-21T21:04:11Z

+        b_good = [1.0, 1.0]
+        prob_good = LinearProblem(A_good, b_good)
+
+        sol_good = solve(prob_good, LUFactorization(); verbose=verbose_all)


Same thing as above here, @test_log (:info ...

jClugstor · 2025-08-21T21:44:45Z

+        prob = LinearProblem(A, b)
+
+        # This should fail due to singularity but not throw
+        sol = solve(prob, LUFactorization(); verbose=verbose)


Suggested change

sol = solve(prob, LUFactorization(); verbose=verbose)

sol = @test_logs (:warn, r"Matrix is singular") match_mode=:any solve(prob, LUFactorization(); verbose=verbose)

jClugstor · 2025-08-21T21:46:32Z

+        b_good = [1.0, 1.0]
+        prob_good = LinearProblem(A_good, b_good)
+
+        sol_good = solve(prob_good, LUFactorization(); verbose=verbose_all)


Suggested change

sol_good = solve(prob_good, LUFactorization(); verbose=verbose_all)

sol_good = @test_logs (:info, r"Operation completed successfully") match_mode=:any solve(prob, LUFactorization(); verbose=verbose)

claude added 5 commits August 18, 2025 19:40

ChrisRackauckas requested a review from jClugstor August 20, 2025 20:37

jClugstor approved these changes Aug 21, 2025

View reviewed changes

jClugstor reviewed Aug 21, 2025

View reviewed changes

ChrisRackauckas closed this Oct 29, 2025

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Uh oh!

Enhanced BLAS/LAPACK return code interpretation and logging#739

Enhanced BLAS/LAPACK return code interpretation and logging#739
ChrisRackauckas-Claude wants to merge 5 commits into
SciML:mainfrom
ChrisRackauckas-Claude:blas-logging-enhancements

ChrisRackauckas-Claude commented Aug 18, 2025

Uh oh!

ChrisRackauckas commented Aug 20, 2025

Uh oh!

jClugstor commented Aug 20, 2025

Uh oh!

ChrisRackauckas commented Aug 21, 2025

Uh oh!

jClugstor left a comment

Uh oh!

jClugstor Aug 21, 2025

Uh oh!

ChrisRackauckas Aug 21, 2025

Uh oh!

jClugstor Aug 21, 2025

Uh oh!

jClugstor Aug 21, 2025 •

edited

Loading

Uh oh!

jClugstor Aug 21, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

	sol = solve(prob, LUFactorization(); verbose=verbose)
	sol = @test_logs (:warn, r"Matrix is singular") match_mode=:any solve(prob, LUFactorization(); verbose=verbose)

	sol_good = solve(prob_good, LUFactorization(); verbose=verbose_all)
	sol_good = @test_logs (:info, r"Operation completed successfully") match_mode=:any solve(prob, LUFactorization(); verbose=verbose)

Uh oh!

Uh oh!

Conversation

ChrisRackauckas-Claude commented Aug 18, 2025

Summary

Changes

1. BLAS/LAPACK Return Code Interpretation

2. Extended Logging Information

3. New Verbosity Controls

4. Integration Example (BLISLUFactorization)

5. Documentation

6. Tests

Example Output

Test Plan

Uh oh!

ChrisRackauckas commented Aug 20, 2025

Uh oh!

jClugstor commented Aug 20, 2025

Uh oh!

ChrisRackauckas commented Aug 21, 2025

Uh oh!

jClugstor left a comment

Choose a reason for hiding this comment

Uh oh!

jClugstor Aug 21, 2025

Choose a reason for hiding this comment

Uh oh!

ChrisRackauckas Aug 21, 2025

Choose a reason for hiding this comment

Uh oh!

jClugstor Aug 21, 2025

Choose a reason for hiding this comment

Uh oh!

jClugstor Aug 21, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jClugstor Aug 21, 2025

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

jClugstor Aug 21, 2025 •

edited

Loading